What signal does an AI detector actually read?
An AI text detector never sees "AI" or "human" as a label in the data — it estimates a probability from two families of signal in the writing itself: how predictable the words are, and how much that predictability changes across the document.
Perplexity
How predictable a piece of text is to a reference language model. When a model can easily guess the next word, perplexity is low. Language models generate by picking statistically likely next tokens, so AI text tends to score consistently low.
Burstiness
How much that predictability varies from sentence to sentence across a document. Human writers vary sentence length and word choice unevenly; models apply similar next-word logic throughout, so their burstiness stays flat.
Why did GPTZero move past perplexity and burstiness?
Perplexity and burstiness are easy to explain but easy to game — paraphrasing tools and "humanizers" exist specifically to push those two numbers toward human-looking ranges. GPTZero says it stopped relying on perplexity and burstiness alone in autumn 2023 and moved to a deep-learning classifier architecture, with perplexity and burstiness retained as just one of several indicators — alongside novel-text search and other model-based signals — rather than the whole verdict. Other detectors, including ZeroGPT, Copyleaks, and Originality.ai, still lean more heavily on perplexity- and burstiness-style scoring.
How accurate are AI detectors, really?
Vendor claims and independent testing tell noticeably different stories, and the gap matters most for anyone facing a consequence — a student, a job applicant, a freelance writer.
| Claim | Source | What it found |
|---|---|---|
| Vendor accuracy claim | Turnitin's July 2026 model update | Under 1% false-positive rate claimed; scores in the 1%-19% range are suppressed and not shown, specifically to avoid flagging borderline text. |
| Independent testing | Stanford-linked research (Liang et al.) | 4-9% false positives on human-written text in broader testing; on TOEFL essays by non-native English writers, over half were misclassified as AI-generated. |
| Character-level tests | Originality.ai | Adding or removing invisible Unicode characters left detection scores essentially unchanged — detectors read words, not bytes. |
The bias finding is the sharpest of the three. In "GPT detectors are biased against non-native English writers" (Liang, Yuksekgonul, Mao, Wu & Zou, published in Patterns, 2023), the authors ran seven widely used detectors against real TOEFL essays and found near-perfect accuracy on native-speaker eighth-grade essays but more than 50% of non-native essays flagged as AI-generated. Their hypothesis: non-native writing tends to have less lexical variety, which lowers perplexity — the exact signal several detectors read as evidence of AI authorship.
What a detector score does — and doesn't — tell you
Even Turnitin, the detector making the strongest accuracy claim, tells institutions not to use its score as the sole basis for an academic-integrity decision. Treat any AI-detection percentage as one input to review, not a verdict — and treat any tool that promises to help you "beat" a detector as making a claim it cannot back up. Detectors score word patterns; they do not read hidden characters, and statistical watermarks like SynthID-Text are verified with a secret key the public tools don't have, not estimated from writing style at all.
This site's own scope is narrower and more honest than a detector or an evasion tool: the AI text cleaner finds and removes character-level artifacts — the kind ChatGPT briefly embedded in 2025 — and shows you exactly what it found with the AI watermark checker. It does not, and cannot, change your detector score.